Collection of tasks for fast prototyping, baselining, finetuning and solving problems with deep learning
Installation • Docs • About • Prediction • Finetuning • Tasks • General Task • Contribute • Community • Website • License
Pip / conda
pip install lightning-flash
Other installations
Pip from source
# with git
pip install git+https://github.com/PytorchLightning/lightning-flash.git@master
# OR from an archive
pip install https://github.com/PyTorchLightning/lightning-flash/archive/master.zip
From source using setuptools
# clone flash repository locally
git clone https://github.com/PyTorchLightning/lightning-flash.git
cd lightning-flash
# install in editable mode
pip install -e .
In case you want to use the extra packages from a specific domain (image, video, text, ...)
pip install "lightning-flash[image]"
See Installation for more options.
Flash is a framework of tasks for fast prototyping, baselining, finetuning and solving business and scientific problems with deep learning. It is focused on:
- Predictions
- Finetuning
- Task-based training
It is built for data scientists, machine learning practitioners, and applied researchers.
Flash is built on top of PyTorch Lightning (by the Lightning team), which is a thin organizational layer on top of PyTorch. If you know PyTorch, you know PyTorch Lightning and Flash already!
As a result, Flash can scale up across any hardware (GPUs, TPUS) with zero changes to your code. It also has the best practices in AI research embedded into each task so you don't have to be a deep learning PhD to leverage its power :)
from flash.text import TranslationTask
# 1. Load finetuned task
model = TranslationTask.load_from_checkpoint("https://flash-weights.s3.amazonaws.com/translation_model_en_ro.pt")
# 2. Translate a few sentences!
predictions = model.predict([
"BBC News went to meet one of the project's first graduates.",
"A recession has come as quickly as 11 months after the first rate hike and as long as 86 months.",
])
print(predictions)
Serve
is a framework agnostic serving engine ! Learn more and check out our examples.
from flash.text import TextClassifier
model = TextClassifier.load_from_checkpoint("https://flash-weights.s3.amazonaws.com/text_classification_model.pt")
model.serve()
Credits to @rlizzo, @hhsecond, @lantiga, @luiscape for building Flash Serve Engine.
First, finetune:
import flash
from flash.core.data.utils import download_data
from flash.image import ImageClassificationData, ImageClassifier
# 1. Download the data
download_data("https://pl-flash-data.s3.amazonaws.com/hymenoptera_data.zip", 'data/')
# 2. Load the data
datamodule = ImageClassificationData.from_folders(
train_folder="data/hymenoptera_data/train/",
val_folder="data/hymenoptera_data/val/",
test_folder="data/hymenoptera_data/test/",
)
# 3. Build the model
model = ImageClassifier(num_classes=datamodule.num_classes, backbone="resnet18")
# 4. Create the trainer. Run once on data
trainer = flash.Trainer(max_epochs=1)
# 5. Finetune the model
trainer.finetune(model, datamodule=datamodule, strategy="freeze")
# 6. Save it!
trainer.save_checkpoint("image_classification_model.pt")
Then use the finetuned model:
from flash.image import ImageClassifier
# load the finetuned model
classifier = ImageClassifier.load_from_checkpoint('image_classification_model.pt')
# predict!
predictions = classifier.predict('data/hymenoptera_data/val/bees/65038344_52a45d090d.jpg')
print(predictions)
Flash is built as a collection of community-built tasks. A task is highly opinionated and laser-focused on solving a single problem well, using state-of-the-art methods.
Flash has an Image Embedder task to encode an image into a vector of image features which can be used for anything like clustering, similarity search or classification.
View example
from flash.core.data.utils import download_data
from flash.image import ImageEmbedder
# 1. Download the data
download_data("https://pl-flash-data.s3.amazonaws.com/hymenoptera_data.zip", 'data/')
# 2. Create an ImageEmbedder with resnet50 trained on imagenet.
embedder = ImageEmbedder(backbone="resnet50", embedding_dim=128)
# 3. Generate an embedding from an image path.
embeddings = embedder.predict('data/hymenoptera_data/predict/153783656_85f9c3ac70.jpg')
# 4. Print embeddings shape
print(embeddings.shape)
Flash has a Summarization task to sum up text from a larger article into a short description.
View example
import flash
from flash.core.data.utils import download_data
from flash.text import SummarizationData, SummarizationTask
# 1. Download the data
download_data("https://pl-flash-data.s3.amazonaws.com/xsum.zip", 'data/')
# 2. Load the data
datamodule = SummarizationData.from_csv(
"input",
"target",
train_file="data/xsum/train.csv",
val_file="data/xsum/valid.csv",
test_file="data/xsum/test.csv",
)
# 3. Build the model
model = SummarizationTask()
# 4. Create the trainer. Run once on data
trainer = flash.Trainer(max_epochs=1, gpus=1, precision=16)
# 5. Fine-tune the model
trainer.finetune(model, datamodule=datamodule)
# 6. Test model
trainer.test()
To run the example:
python flash_examples/finetuning/summarization.py
Flash has a Tabular Classification task to tackle any tabular classification problem.
View example
To illustrate, say we want to build a model to predict if a passenger survived on the Titanic.
from torchmetrics.classification import Accuracy, Precision, Recall
import flash
from flash.core.data.utils import download_data
from flash.tabular import TabularClassifier, TabularClassificationData
# 1. Download the data
download_data("https://pl-flash-data.s3.amazonaws.com/titanic.zip", 'data/')
# 2. Load the data
datamodule = TabularClassificationData.from_csv(
["Sex", "Age", "SibSp", "Parch", "Ticket", "Cabin", "Embarked"],
"Fare",
target_fields="Survived",
train_file="./data/titanic/titanic.csv",
test_file="./data/titanic/test.csv",
val_split=0.25,
)
# 3. Build the model
model = TabularClassifier.from_data(datamodule, metrics=[Accuracy(), Precision(), Recall()])
# 4. Create the trainer. Run 10 times on data
trainer = flash.Trainer(max_epochs=10)
# 5. Train the model
trainer.fit(model, datamodule=datamodule)
# 6. Test model
trainer.test()
# 7. Predict!
predictions = model.predict("data/titanic/titanic.csv")
print(predictions)
To run the example:
python flash_examples/finetuning/tabular_data.py
Flash has an Object Detection task to identify and locate objects in images.
View example
To illustrate, say we want to build a model on a tiny coco dataset.
import flash
from flash.core.data.utils import download_data
from flash.image import ObjectDetectionData, ObjectDetector
# 1. Download the data
# Dataset Credit: https://www.kaggle.com/ultralytics/coco128
download_data("https://github.com/zhiqwang/yolov5-rt-stack/releases/download/v0.3.0/coco128.zip", "data/")
# 2. Load the Data
datamodule = ObjectDetectionData.from_coco(
train_folder="data/coco128/images/train2017/",
train_ann_file="data/coco128/annotations/instances_train2017.json",
batch_size=2
)
# 3. Build the model
model = ObjectDetector(num_classes=datamodule.num_classes)
# 4. Create the trainer. Run twice on data
trainer = flash.Trainer(max_epochs=3)
# 5. Finetune the model
trainer.fit(model, datamodule=datamodule)
# 6. Save it!
trainer.save_checkpoint("object_detection_model.pt")
To run the example:
python flash_examples/finetuning/object_detection.py
Flash has a Video Classification task to classify videos using PyTorchVideo.
View example
To illustrate, say we want to build a model to classify the kinetics data set.
import os
from torch.utils.data.sampler import RandomSampler
import flash
from flash.core.data.utils import download_data
from flash.video import VideoClassificationData, VideoClassifier
# 1. Download a video clip dataset. Find more datasets at https://pytorchvideo.readthedocs.io/en/latest/data.html
download_data("https://pl-flash-data.s3.amazonaws.com/kinetics.zip")
# 2. Load the Data
datamodule = VideoClassificationData.from_folders(
train_folder=os.path.join(flash.PROJECT_ROOT, "data/kinetics/train"),
val_folder=os.path.join(flash.PROJECT_ROOT, "data/kinetics/val"),
predict_folder=os.path.join(flash.PROJECT_ROOT, "data/kinetics/predict"),
batch_size=8,
clip_sampler="uniform",
clip_duration=1,
video_sampler=RandomSampler,
decode_audio=False,
num_workers=8,
)
# 3. Build the model
model = VideoClassifier(
backbone="x3d_xs", num_classes=datamodule.num_classes, pretrained=False
)
# 4. Create the trainer
trainer = flash.Trainer(max_epochs=3)
# 5. Finetune the model
trainer.finetune(model, datamodule=datamodule)
# 6. Save it!
trainer.save_checkpoint("video_classification.pt")
To run the example:
python flash_examples/finetuning/video_classification.py
Flash has a Semantic Segmentation task for segmentation of images.
View example
To illustrate, say we want to finetune a model on this data from the Lyft Udacity Challenge.
import flash
from flash.core.data.utils import download_data
from flash.image import SemanticSegmentation, SemanticSegmentationData
# 1. Download the Data
download_data("https://github.com/ongchinkiat/LyftPerceptionChallenge/releases/download/v0.1/carla-capture-20180513A.zip", "data/")
# 2. Load the Data
datamodule = SemanticSegmentationData.from_folders(
train_folder="data/CameraRGB",
train_target_folder="data/CameraSeg",
batch_size=4,
val_split=0.3,
image_size=(200, 200),
num_classes=21,
)
# 3. Build the model
model = SemanticSegmentation(
backbone="torchvision/fcn_resnet50",
num_classes=datamodule.num_classes,
)
# 4. Create the trainer
trainer = flash.Trainer(max_epochs=3)
# 5. Finetune the model
trainer.finetune(model, datamodule=datamodule)
# 6. Save it!
trainer.save_checkpoint("semantic_segmentation_model.pt")
To run the example:
python flash_examples/finetuning/semantic_segmentation.py
Flash has a Style Transfer task for Neural Style Transfer (NST) with Pystiche.
View example
To illustrate, say we want to train an NST model to transfer the style from the paint demo image to the COCO data set.
import pystiche.demo
import flash
from flash.core.data.utils import download_data
from flash.image.style_transfer import StyleTransfer, StyleTransferData
# 1. Download the Data
download_data("https://github.com/zhiqwang/yolov5-rt-stack/releases/download/v0.3.0/coco128.zip", "data/")
# 2. Load the Data
datamodule = StyleTransferData.from_folders(train_folder="data/coco128/images", batch_size=4)
# 3. Load the style image
style_image = pystiche.demo.images()["paint"].read(size=256)
# 4. Build the model
model = StyleTransfer(style_image)
# 5. Create the trainer
trainer = flash.Trainer(max_epochs=2)
# 6. Train the model
trainer.fit(model, datamodule=datamodule)
# 7. Save it!
trainer.save_checkpoint("style_transfer_model.pt")
To run the example:
python flash_examples/finetuning/style_transfer.py
Flash comes prebuilt with a task to handle a huge portion of deep learning problems.
import flash
from torch import nn, optim
from torch.utils.data import DataLoader, random_split
from torchvision import transforms, datasets
# model
model = nn.Sequential(
nn.Flatten(),
nn.Linear(28 * 28, 128),
nn.ReLU(),
nn.Linear(128, 10)
)
# data
dataset = datasets.MNIST('./data_folder', download=True, transform=transforms.ToTensor())
train, val = random_split(dataset, [55000, 5000])
# task
classifier = flash.Task(model, loss_fn=nn.functional.cross_entropy, optimizer=optim.Adam)
# train
flash.Trainer().fit(classifier, DataLoader(train), DataLoader(val))
Tasks can be built in just a few minutes because Flash is built on top of PyTorch Lightning LightningModules, which are infinitely extensible and let you train across GPUs, TPUs etc without doing any code changes.
import torch
import torch.nn.functional as F
from torchmetrics import Accuracy
from typing import Callable, Mapping, Sequence, Type, Union
from flash.core.classification import ClassificationTask
class LinearClassifier(ClassificationTask):
def __init__(
self,
num_inputs,
num_classes,
loss_fn: Callable = F.cross_entropy,
optimizer: Type[torch.optim.Optimizer] = torch.optim.SGD,
metrics: Union[Callable, Mapping, Sequence, None] = [Accuracy()],
learning_rate: float = 1e-3,
):
super().__init__(
model=None,
loss_fn=loss_fn,
optimizer=optimizer,
metrics=metrics,
learning_rate=learning_rate,
)
self.save_hyperparameters()
self.linear = torch.nn.Linear(num_inputs, num_classes)
def forward(self, x):
return self.linear(x)
classifier = LinearClassifier(128, 10)
...
When you reach the limits of the flexibility provided by Flash, then seamlessly transition to PyTorch Lightning which gives you the most flexibility because it is simply organized PyTorch.
Predictions from image and video tasks can be visualized through an integration with FiftyOne, allowing you to better understand and analyze how your model is performing.
from flash.core.data.utils import download_data
from flash.core.integrations.fiftyone import visualize
from flash.image import ObjectDetector
from flash.image.detection.serialization import FiftyOneDetectionLabels
# 1. Download the data
# Dataset Credit: https://www.kaggle.com/ultralytics/coco128
download_data(
"https://github.com/zhiqwang/yolov5-rt-stack/releases/download/v0.3.0/coco128.zip",
"data/",
)
# 2. Load the model from a checkpoint and use the FiftyOne serializer
model = ObjectDetector.load_from_checkpoint(
"https://flash-weights.s3.amazonaws.com/object_detection_model.pt"
)
model.serializer = FiftyOneDetectionLabels()
# 3. Detect the object on the images
filepaths = [
"data/coco128/images/train2017/000000000025.jpg",
"data/coco128/images/train2017/000000000520.jpg",
"data/coco128/images/train2017/000000000532.jpg",
]
predictions = model.predict(filepaths)
# 4. Visualize predictions in FiftyOne App
session = visualize(predictions, filepaths=filepaths)
The lightning + Flash team is hard at work building more tasks for common deep-learning use cases. But we're looking for incredible contributors like you to submit new tasks!
Join our Slack and/or read our CONTRIBUTING guidelines to get help becoming a contributor!
For help or questions, join our huge community on Slack!
We’re excited to continue the strong legacy of opensource software and have been inspired over the years by Caffee, Theano, Keras, PyTorch, torchbearer, and fast.ai. When/if a paper is written about this, we’ll be happy to cite these frameworks and the corresponding authors.
Flash leverages models from torchvision, huggingface/transformers, timm, open3d-ml for pointcloud, pytorch-tabnet, and asteroid for the vision
, text
, tabular
, and audio
tasks respectively. Also supports self-supervised backbones from bolts.
Please observe the Apache 2.0 license that is listed in this repository. In addition the Lightning framework is Patent Pending.